xref: /freebsd/sys/kern/kern_racct.c (revision 5d1fba01ca07030a013815746c2c056b5c4a8a49)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2010 The FreeBSD Foundation
5  *
6  * This software was developed by Edward Tomasz Napierala under sponsorship
7  * from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 #include "opt_sched.h"
33 
34 #include <sys/param.h>
35 #include <sys/buf.h>
36 #include <sys/systm.h>
37 #include <sys/eventhandler.h>
38 #include <sys/jail.h>
39 #include <sys/kernel.h>
40 #include <sys/kthread.h>
41 #include <sys/lock.h>
42 #include <sys/loginclass.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
45 #include <sys/proc.h>
46 #include <sys/racct.h>
47 #include <sys/resourcevar.h>
48 #include <sys/sbuf.h>
49 #include <sys/sched.h>
50 #include <sys/sdt.h>
51 #include <sys/smp.h>
52 #include <sys/sx.h>
53 #include <sys/sysctl.h>
54 #include <sys/sysproto.h>
55 #include <sys/umtxvar.h>
56 #include <machine/smp.h>
57 
58 #ifdef RCTL
59 #include <sys/rctl.h>
60 #endif
61 
62 FEATURE(racct, "Resource Accounting");
63 
64 /*
65  * Do not block processes that have their %cpu usage <= pcpu_threshold.
66  */
67 static int pcpu_threshold = 1;
68 #ifdef RACCT_DEFAULT_TO_DISABLED
69 bool __read_frequently racct_enable = false;
70 #else
71 bool __read_frequently racct_enable = true;
72 #endif
73 
74 SYSCTL_NODE(_kern, OID_AUTO, racct, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
75     "Resource Accounting");
76 SYSCTL_BOOL(_kern_racct, OID_AUTO, enable, CTLFLAG_RDTUN, &racct_enable,
77     0, "Enable RACCT/RCTL");
78 SYSCTL_UINT(_kern_racct, OID_AUTO, pcpu_threshold, CTLFLAG_RW, &pcpu_threshold,
79     0, "Processes with higher %cpu usage than this value can be throttled.");
80 
81 /*
82  * How many seconds it takes to use the scheduler %cpu calculations.  When a
83  * process starts, we compute its %cpu usage by dividing its runtime by the
84  * process wall clock time.  After RACCT_PCPU_SECS pass, we use the value
85  * provided by the scheduler.
86  */
87 #define RACCT_PCPU_SECS		3
88 
89 struct mtx racct_lock;
90 MTX_SYSINIT(racct_lock, &racct_lock, "racct lock", MTX_DEF);
91 
92 static uma_zone_t racct_zone;
93 
94 static void racct_sub_racct(struct racct *dest, const struct racct *src);
95 static void racct_sub_cred_locked(struct ucred *cred, int resource,
96 		uint64_t amount);
97 static void racct_add_cred_locked(struct ucred *cred, int resource,
98 		uint64_t amount);
99 
100 SDT_PROVIDER_DEFINE(racct);
101 SDT_PROBE_DEFINE3(racct, , rusage, add,
102     "struct proc *", "int", "uint64_t");
103 SDT_PROBE_DEFINE3(racct, , rusage, add__failure,
104     "struct proc *", "int", "uint64_t");
105 SDT_PROBE_DEFINE3(racct, , rusage, add__buf,
106     "struct proc *", "const struct buf *", "int");
107 SDT_PROBE_DEFINE3(racct, , rusage, add__cred,
108     "struct ucred *", "int", "uint64_t");
109 SDT_PROBE_DEFINE3(racct, , rusage, add__force,
110     "struct proc *", "int", "uint64_t");
111 SDT_PROBE_DEFINE3(racct, , rusage, set,
112     "struct proc *", "int", "uint64_t");
113 SDT_PROBE_DEFINE3(racct, , rusage, set__failure,
114     "struct proc *", "int", "uint64_t");
115 SDT_PROBE_DEFINE3(racct, , rusage, set__force,
116     "struct proc *", "int", "uint64_t");
117 SDT_PROBE_DEFINE3(racct, , rusage, sub,
118     "struct proc *", "int", "uint64_t");
119 SDT_PROBE_DEFINE3(racct, , rusage, sub__cred,
120     "struct ucred *", "int", "uint64_t");
121 SDT_PROBE_DEFINE1(racct, , racct, create,
122     "struct racct *");
123 SDT_PROBE_DEFINE1(racct, , racct, destroy,
124     "struct racct *");
125 SDT_PROBE_DEFINE2(racct, , racct, join,
126     "struct racct *", "struct racct *");
127 SDT_PROBE_DEFINE2(racct, , racct, join__failure,
128     "struct racct *", "struct racct *");
129 SDT_PROBE_DEFINE2(racct, , racct, leave,
130     "struct racct *", "struct racct *");
131 
132 int racct_types[] = {
133 	[RACCT_CPU] =
134 		RACCT_IN_MILLIONS,
135 	[RACCT_DATA] =
136 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
137 	[RACCT_STACK] =
138 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
139 	[RACCT_CORE] =
140 		RACCT_DENIABLE,
141 	[RACCT_RSS] =
142 		RACCT_RECLAIMABLE,
143 	[RACCT_MEMLOCK] =
144 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
145 	[RACCT_NPROC] =
146 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
147 	[RACCT_NOFILE] =
148 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
149 	[RACCT_VMEM] =
150 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
151 	[RACCT_NPTS] =
152 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
153 	[RACCT_SWAP] =
154 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
155 	[RACCT_NTHR] =
156 		RACCT_RECLAIMABLE | RACCT_DENIABLE,
157 	[RACCT_MSGQQUEUED] =
158 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
159 	[RACCT_MSGQSIZE] =
160 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
161 	[RACCT_NMSGQ] =
162 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
163 	[RACCT_NSEM] =
164 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
165 	[RACCT_NSEMOP] =
166 		RACCT_RECLAIMABLE | RACCT_INHERITABLE | RACCT_DENIABLE,
167 	[RACCT_NSHM] =
168 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
169 	[RACCT_SHMSIZE] =
170 		RACCT_RECLAIMABLE | RACCT_DENIABLE | RACCT_SLOPPY,
171 	[RACCT_WALLCLOCK] =
172 		RACCT_IN_MILLIONS,
173 	[RACCT_PCTCPU] =
174 		RACCT_DECAYING | RACCT_DENIABLE | RACCT_IN_MILLIONS,
175 	[RACCT_READBPS] =
176 		RACCT_DECAYING,
177 	[RACCT_WRITEBPS] =
178 		RACCT_DECAYING,
179 	[RACCT_READIOPS] =
180 		RACCT_DECAYING,
181 	[RACCT_WRITEIOPS] =
182 		RACCT_DECAYING };
183 
184 static const fixpt_t RACCT_DECAY_FACTOR = 0.3 * FSCALE;
185 
186 #ifdef SCHED_4BSD
187 /*
188  * Contains intermediate values for %cpu calculations to avoid using floating
189  * point in the kernel.
190  * ccpu_exp[k] = FSCALE * (ccpu/FSCALE)^k = FSCALE * exp(-k/20)
191  * It is needed only for the 4BSD scheduler, because in ULE, the ccpu equals to
192  * zero so the calculations are more straightforward.
193  */
194 fixpt_t ccpu_exp[] = {
195 	[0] = FSCALE * 1,
196 	[1] = FSCALE * 0.95122942450071400909,
197 	[2] = FSCALE * 0.90483741803595957316,
198 	[3] = FSCALE * 0.86070797642505780722,
199 	[4] = FSCALE * 0.81873075307798185866,
200 	[5] = FSCALE * 0.77880078307140486824,
201 	[6] = FSCALE * 0.74081822068171786606,
202 	[7] = FSCALE * 0.70468808971871343435,
203 	[8] = FSCALE * 0.67032004603563930074,
204 	[9] = FSCALE * 0.63762815162177329314,
205 	[10] = FSCALE * 0.60653065971263342360,
206 	[11] = FSCALE * 0.57694981038048669531,
207 	[12] = FSCALE * 0.54881163609402643262,
208 	[13] = FSCALE * 0.52204577676101604789,
209 	[14] = FSCALE * 0.49658530379140951470,
210 	[15] = FSCALE * 0.47236655274101470713,
211 	[16] = FSCALE * 0.44932896411722159143,
212 	[17] = FSCALE * 0.42741493194872666992,
213 	[18] = FSCALE * 0.40656965974059911188,
214 	[19] = FSCALE * 0.38674102345450120691,
215 	[20] = FSCALE * 0.36787944117144232159,
216 	[21] = FSCALE * 0.34993774911115535467,
217 	[22] = FSCALE * 0.33287108369807955328,
218 	[23] = FSCALE * 0.31663676937905321821,
219 	[24] = FSCALE * 0.30119421191220209664,
220 	[25] = FSCALE * 0.28650479686019010032,
221 	[26] = FSCALE * 0.27253179303401260312,
222 	[27] = FSCALE * 0.25924026064589150757,
223 	[28] = FSCALE * 0.24659696394160647693,
224 	[29] = FSCALE * 0.23457028809379765313,
225 	[30] = FSCALE * 0.22313016014842982893,
226 	[31] = FSCALE * 0.21224797382674305771,
227 	[32] = FSCALE * 0.20189651799465540848,
228 	[33] = FSCALE * 0.19204990862075411423,
229 	[34] = FSCALE * 0.18268352405273465022,
230 	[35] = FSCALE * 0.17377394345044512668,
231 	[36] = FSCALE * 0.16529888822158653829,
232 	[37] = FSCALE * 0.15723716631362761621,
233 	[38] = FSCALE * 0.14956861922263505264,
234 	[39] = FSCALE * 0.14227407158651357185,
235 	[40] = FSCALE * 0.13533528323661269189,
236 	[41] = FSCALE * 0.12873490358780421886,
237 	[42] = FSCALE * 0.12245642825298191021,
238 	[43] = FSCALE * 0.11648415777349695786,
239 	[44] = FSCALE * 0.11080315836233388333,
240 	[45] = FSCALE * 0.10539922456186433678,
241 	[46] = FSCALE * 0.10025884372280373372,
242 	[47] = FSCALE * 0.09536916221554961888,
243 	[48] = FSCALE * 0.09071795328941250337,
244 	[49] = FSCALE * 0.08629358649937051097,
245 	[50] = FSCALE * 0.08208499862389879516,
246 	[51] = FSCALE * 0.07808166600115315231,
247 	[52] = FSCALE * 0.07427357821433388042,
248 	[53] = FSCALE * 0.07065121306042958674,
249 	[54] = FSCALE * 0.06720551273974976512,
250 	[55] = FSCALE * 0.06392786120670757270,
251 	[56] = FSCALE * 0.06081006262521796499,
252 	[57] = FSCALE * 0.05784432087483846296,
253 	[58] = FSCALE * 0.05502322005640722902,
254 	[59] = FSCALE * 0.05233970594843239308,
255 	[60] = FSCALE * 0.04978706836786394297,
256 	[61] = FSCALE * 0.04735892439114092119,
257 	[62] = FSCALE * 0.04504920239355780606,
258 	[63] = FSCALE * 0.04285212686704017991,
259 	[64] = FSCALE * 0.04076220397836621516,
260 	[65] = FSCALE * 0.03877420783172200988,
261 	[66] = FSCALE * 0.03688316740124000544,
262 	[67] = FSCALE * 0.03508435410084502588,
263 	[68] = FSCALE * 0.03337326996032607948,
264 	[69] = FSCALE * 0.03174563637806794323,
265 	[70] = FSCALE * 0.03019738342231850073,
266 	[71] = FSCALE * 0.02872463965423942912,
267 	[72] = FSCALE * 0.02732372244729256080,
268 	[73] = FSCALE * 0.02599112877875534358,
269 	[74] = FSCALE * 0.02472352647033939120,
270 	[75] = FSCALE * 0.02351774585600910823,
271 	[76] = FSCALE * 0.02237077185616559577,
272 	[77] = FSCALE * 0.02127973643837716938,
273 	[78] = FSCALE * 0.02024191144580438847,
274 	[79] = FSCALE * 0.01925470177538692429,
275 	[80] = FSCALE * 0.01831563888873418029,
276 	[81] = FSCALE * 0.01742237463949351138,
277 	[82] = FSCALE * 0.01657267540176124754,
278 	[83] = FSCALE * 0.01576441648485449082,
279 	[84] = FSCALE * 0.01499557682047770621,
280 	[85] = FSCALE * 0.01426423390899925527,
281 	[86] = FSCALE * 0.01356855901220093175,
282 	[87] = FSCALE * 0.01290681258047986886,
283 	[88] = FSCALE * 0.01227733990306844117,
284 	[89] = FSCALE * 0.01167856697039544521,
285 	[90] = FSCALE * 0.01110899653824230649,
286 	[91] = FSCALE * 0.01056720438385265337,
287 	[92] = FSCALE * 0.01005183574463358164,
288 	[93] = FSCALE * 0.00956160193054350793,
289 	[94] = FSCALE * 0.00909527710169581709,
290 	[95] = FSCALE * 0.00865169520312063417,
291 	[96] = FSCALE * 0.00822974704902002884,
292 	[97] = FSCALE * 0.00782837754922577143,
293 	[98] = FSCALE * 0.00744658307092434051,
294 	[99] = FSCALE * 0.00708340892905212004,
295 	[100] = FSCALE * 0.00673794699908546709,
296 	[101] = FSCALE * 0.00640933344625638184,
297 	[102] = FSCALE * 0.00609674656551563610,
298 	[103] = FSCALE * 0.00579940472684214321,
299 	[104] = FSCALE * 0.00551656442076077241,
300 	[105] = FSCALE * 0.00524751839918138427,
301 	[106] = FSCALE * 0.00499159390691021621,
302 	[107] = FSCALE * 0.00474815099941147558,
303 	[108] = FSCALE * 0.00451658094261266798,
304 	[109] = FSCALE * 0.00429630469075234057,
305 	[110] = FSCALE * 0.00408677143846406699,
306 };
307 #endif
308 
309 #define	CCPU_EXP_MAX	110
310 
311 /*
312  * This function is analogical to the getpcpu() function in the ps(1) command.
313  * They should both calculate in the same way so that the racct %cpu
314  * calculations are consistent with the values shown by the ps(1) tool.
315  * The calculations are more complex in the 4BSD scheduler because of the value
316  * of the ccpu variable.  In ULE it is defined to be zero which saves us some
317  * work.
318  */
319 static uint64_t
racct_getpcpu(struct proc * p,u_int pcpu)320 racct_getpcpu(struct proc *p, u_int pcpu)
321 {
322 	u_int swtime;
323 #ifdef SCHED_4BSD
324 	fixpt_t pctcpu, pctcpu_next;
325 #endif
326 	fixpt_t p_pctcpu;
327 	struct thread *td;
328 
329 	ASSERT_RACCT_ENABLED();
330 	KASSERT((p->p_flag & P_IDLEPROC) == 0,
331 	    ("racct_getpcpu: idle process %p", p));
332 
333 	swtime = (ticks - p->p_swtick) / hz;
334 
335 	/*
336 	 * For short-lived processes, the sched_pctcpu() returns small
337 	 * values even for cpu intensive processes.  Therefore we use
338 	 * our own estimate in this case.
339 	 */
340 	if (swtime < RACCT_PCPU_SECS)
341 		return (pcpu);
342 
343 	p_pctcpu = 0;
344 	FOREACH_THREAD_IN_PROC(p, td) {
345 		thread_lock(td);
346 #ifdef SCHED_4BSD
347 		pctcpu = sched_pctcpu(td);
348 		/* Count also the yet unfinished second. */
349 		pctcpu_next = (pctcpu * ccpu_exp[1]) >> FSHIFT;
350 		pctcpu_next += sched_pctcpu_delta(td);
351 		p_pctcpu += max(pctcpu, pctcpu_next);
352 #else
353 		/*
354 		 * In ULE the %cpu statistics are updated on every
355 		 * sched_pctcpu() call.  So special calculations to
356 		 * account for the latest (unfinished) second are
357 		 * not needed.
358 		 */
359 		p_pctcpu += sched_pctcpu(td);
360 #endif
361 		thread_unlock(td);
362 	}
363 
364 #ifdef SCHED_4BSD
365 	if (swtime <= CCPU_EXP_MAX)
366 		return ((100 * (uint64_t)p_pctcpu * 1000000) /
367 		    (FSCALE - ccpu_exp[swtime]));
368 #endif
369 
370 	return ((100 * (uint64_t)p_pctcpu * 1000000) / FSCALE);
371 }
372 
373 static void
racct_add_racct(struct racct * dest,const struct racct * src)374 racct_add_racct(struct racct *dest, const struct racct *src)
375 {
376 	int i;
377 
378 	ASSERT_RACCT_ENABLED();
379 	RACCT_LOCK_ASSERT();
380 
381 	/*
382 	 * Update resource usage in dest.
383 	 */
384 	for (i = 0; i <= RACCT_MAX; i++) {
385 		KASSERT(dest->r_resources[i] >= 0,
386 		    ("%s: resource %d propagation meltdown: dest < 0",
387 		    __func__, i));
388 		KASSERT(src->r_resources[i] >= 0,
389 		    ("%s: resource %d propagation meltdown: src < 0",
390 		    __func__, i));
391 		dest->r_resources[i] += src->r_resources[i];
392 	}
393 }
394 
395 static void
racct_sub_racct(struct racct * dest,const struct racct * src)396 racct_sub_racct(struct racct *dest, const struct racct *src)
397 {
398 	int i;
399 
400 	ASSERT_RACCT_ENABLED();
401 	RACCT_LOCK_ASSERT();
402 
403 	/*
404 	 * Update resource usage in dest.
405 	 */
406 	for (i = 0; i <= RACCT_MAX; i++) {
407 		if (!RACCT_IS_SLOPPY(i) && !RACCT_IS_DECAYING(i)) {
408 			KASSERT(dest->r_resources[i] >= 0,
409 			    ("%s: resource %d propagation meltdown: dest < 0",
410 			    __func__, i));
411 			KASSERT(src->r_resources[i] >= 0,
412 			    ("%s: resource %d propagation meltdown: src < 0",
413 			    __func__, i));
414 			KASSERT(src->r_resources[i] <= dest->r_resources[i],
415 			    ("%s: resource %d propagation meltdown: src > dest",
416 			    __func__, i));
417 		}
418 		if (RACCT_CAN_DROP(i)) {
419 			dest->r_resources[i] -= src->r_resources[i];
420 			if (dest->r_resources[i] < 0)
421 				dest->r_resources[i] = 0;
422 		}
423 	}
424 }
425 
426 void
racct_create(struct racct ** racctp)427 racct_create(struct racct **racctp)
428 {
429 
430 	if (!racct_enable)
431 		return;
432 
433 	SDT_PROBE1(racct, , racct, create, racctp);
434 
435 	KASSERT(*racctp == NULL, ("racct already allocated"));
436 
437 	*racctp = uma_zalloc(racct_zone, M_WAITOK | M_ZERO);
438 }
439 
440 static void
racct_destroy_locked(struct racct ** racctp)441 racct_destroy_locked(struct racct **racctp)
442 {
443 	struct racct *racct;
444 	int i;
445 
446 	ASSERT_RACCT_ENABLED();
447 
448 	SDT_PROBE1(racct, , racct, destroy, racctp);
449 
450 	RACCT_LOCK_ASSERT();
451 	KASSERT(racctp != NULL, ("NULL racctp"));
452 	KASSERT(*racctp != NULL, ("NULL racct"));
453 
454 	racct = *racctp;
455 
456 	for (i = 0; i <= RACCT_MAX; i++) {
457 		if (RACCT_IS_SLOPPY(i))
458 			continue;
459 		if (!RACCT_IS_RECLAIMABLE(i))
460 			continue;
461 		KASSERT(racct->r_resources[i] == 0,
462 		    ("destroying non-empty racct: "
463 		    "%ju allocated for resource %d\n",
464 		    racct->r_resources[i], i));
465 	}
466 	uma_zfree(racct_zone, racct);
467 	*racctp = NULL;
468 }
469 
470 void
racct_destroy(struct racct ** racct)471 racct_destroy(struct racct **racct)
472 {
473 
474 	if (!racct_enable)
475 		return;
476 
477 	RACCT_LOCK();
478 	racct_destroy_locked(racct);
479 	RACCT_UNLOCK();
480 }
481 
482 /*
483  * Increase consumption of 'resource' by 'amount' for 'racct',
484  * but not its parents.  Differently from other cases, 'amount' here
485  * may be less than zero.
486  */
487 static void
racct_adjust_resource(struct racct * racct,int resource,int64_t amount)488 racct_adjust_resource(struct racct *racct, int resource,
489     int64_t amount)
490 {
491 
492 	ASSERT_RACCT_ENABLED();
493 	RACCT_LOCK_ASSERT();
494 	KASSERT(racct != NULL, ("NULL racct"));
495 
496 	racct->r_resources[resource] += amount;
497 	if (racct->r_resources[resource] < 0) {
498 		KASSERT(RACCT_IS_SLOPPY(resource) || RACCT_IS_DECAYING(resource),
499 		    ("%s: resource %d usage < 0", __func__, resource));
500 		racct->r_resources[resource] = 0;
501 	}
502 
503 	/*
504 	 * There are some cases where the racct %cpu resource would grow
505 	 * beyond 100% per core.  For example in racct_proc_exit() we add
506 	 * the process %cpu usage to the ucred racct containers.  If too
507 	 * many processes terminated in a short time span, the ucred %cpu
508 	 * resource could grow too much.  Also, the 4BSD scheduler sometimes
509 	 * returns for a thread more than 100% cpu usage. So we set a sane
510 	 * boundary here to 100% * the maximum number of CPUs.
511 	 */
512 	if ((resource == RACCT_PCTCPU) &&
513 	    (racct->r_resources[RACCT_PCTCPU] > 100 * 1000000 * (int64_t)MAXCPU))
514 		racct->r_resources[RACCT_PCTCPU] = 100 * 1000000 * (int64_t)MAXCPU;
515 }
516 
517 static int
racct_add_locked(struct proc * p,int resource,uint64_t amount,int force)518 racct_add_locked(struct proc *p, int resource, uint64_t amount, int force)
519 {
520 #ifdef RCTL
521 	int error;
522 #endif
523 
524 	ASSERT_RACCT_ENABLED();
525 
526 	/*
527 	 * We need proc lock to dereference p->p_ucred.
528 	 */
529 	PROC_LOCK_ASSERT(p, MA_OWNED);
530 
531 #ifdef RCTL
532 	error = rctl_enforce(p, resource, amount);
533 	if (error && !force && RACCT_IS_DENIABLE(resource)) {
534 		SDT_PROBE3(racct, , rusage, add__failure, p, resource, amount);
535 		return (error);
536 	}
537 #endif
538 	racct_adjust_resource(p->p_racct, resource, amount);
539 	racct_add_cred_locked(p->p_ucred, resource, amount);
540 
541 	return (0);
542 }
543 
544 /*
545  * Increase allocation of 'resource' by 'amount' for process 'p'.
546  * Return 0 if it's below limits, or errno, if it's not.
547  */
548 int
racct_add(struct proc * p,int resource,uint64_t amount)549 racct_add(struct proc *p, int resource, uint64_t amount)
550 {
551 	int error;
552 
553 	if (!racct_enable)
554 		return (0);
555 
556 	SDT_PROBE3(racct, , rusage, add, p, resource, amount);
557 
558 	RACCT_LOCK();
559 	error = racct_add_locked(p, resource, amount, 0);
560 	RACCT_UNLOCK();
561 	return (error);
562 }
563 
564 /*
565  * Increase allocation of 'resource' by 'amount' for process 'p'.
566  * Doesn't check for limits and never fails.
567  */
568 void
racct_add_force(struct proc * p,int resource,uint64_t amount)569 racct_add_force(struct proc *p, int resource, uint64_t amount)
570 {
571 
572 	if (!racct_enable)
573 		return;
574 
575 	SDT_PROBE3(racct, , rusage, add__force, p, resource, amount);
576 
577 	RACCT_LOCK();
578 	racct_add_locked(p, resource, amount, 1);
579 	RACCT_UNLOCK();
580 }
581 
582 static void
racct_add_cred_locked(struct ucred * cred,int resource,uint64_t amount)583 racct_add_cred_locked(struct ucred *cred, int resource, uint64_t amount)
584 {
585 	struct prison *pr;
586 
587 	ASSERT_RACCT_ENABLED();
588 
589 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, amount);
590 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
591 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
592 		    amount);
593 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, amount);
594 }
595 
596 /*
597  * Increase allocation of 'resource' by 'amount' for credential 'cred'.
598  * Doesn't check for limits and never fails.
599  */
600 void
racct_add_cred(struct ucred * cred,int resource,uint64_t amount)601 racct_add_cred(struct ucred *cred, int resource, uint64_t amount)
602 {
603 
604 	if (!racct_enable)
605 		return;
606 
607 	SDT_PROBE3(racct, , rusage, add__cred, cred, resource, amount);
608 
609 	RACCT_LOCK();
610 	racct_add_cred_locked(cred, resource, amount);
611 	RACCT_UNLOCK();
612 }
613 
614 /*
615  * Account for disk IO resource consumption.  Checks for limits,
616  * but never fails, due to disk limits being undeniable.
617  */
618 void
racct_add_buf(struct proc * p,const struct buf * bp,int is_write)619 racct_add_buf(struct proc *p, const struct buf *bp, int is_write)
620 {
621 
622 	ASSERT_RACCT_ENABLED();
623 	PROC_LOCK_ASSERT(p, MA_OWNED);
624 
625 	SDT_PROBE3(racct, , rusage, add__buf, p, bp, is_write);
626 
627 	RACCT_LOCK();
628 	if (is_write) {
629 		racct_add_locked(curproc, RACCT_WRITEBPS, bp->b_bcount, 1);
630 		racct_add_locked(curproc, RACCT_WRITEIOPS, 1, 1);
631 	} else {
632 		racct_add_locked(curproc, RACCT_READBPS, bp->b_bcount, 1);
633 		racct_add_locked(curproc, RACCT_READIOPS, 1, 1);
634 	}
635 	RACCT_UNLOCK();
636 }
637 
638 static int
racct_set_locked(struct proc * p,int resource,uint64_t amount,int force)639 racct_set_locked(struct proc *p, int resource, uint64_t amount, int force)
640 {
641 	int64_t old_amount, decayed_amount, diff_proc, diff_cred;
642 #ifdef RCTL
643 	int error;
644 #endif
645 
646 	ASSERT_RACCT_ENABLED();
647 
648 	/*
649 	 * We need proc lock to dereference p->p_ucred.
650 	 */
651 	PROC_LOCK_ASSERT(p, MA_OWNED);
652 
653 	old_amount = p->p_racct->r_resources[resource];
654 	/*
655 	 * The diffs may be negative.
656 	 */
657 	diff_proc = amount - old_amount;
658 	if (resource == RACCT_PCTCPU) {
659 		/*
660 		 * Resources in per-credential racct containers may decay.
661 		 * If this is the case, we need to calculate the difference
662 		 * between the new amount and the proportional value of the
663 		 * old amount that has decayed in the ucred racct containers.
664 		 */
665 		decayed_amount = old_amount * RACCT_DECAY_FACTOR / FSCALE;
666 		diff_cred = amount - decayed_amount;
667 	} else
668 		diff_cred = diff_proc;
669 #ifdef notyet
670 	KASSERT(diff_proc >= 0 || RACCT_CAN_DROP(resource),
671 	    ("%s: usage of non-droppable resource %d dropping", __func__,
672 	     resource));
673 #endif
674 #ifdef RCTL
675 	if (diff_proc > 0) {
676 		error = rctl_enforce(p, resource, diff_proc);
677 		if (error && !force && RACCT_IS_DENIABLE(resource)) {
678 			SDT_PROBE3(racct, , rusage, set__failure, p, resource,
679 			    amount);
680 			return (error);
681 		}
682 	}
683 #endif
684 	racct_adjust_resource(p->p_racct, resource, diff_proc);
685 	if (diff_cred > 0)
686 		racct_add_cred_locked(p->p_ucred, resource, diff_cred);
687 	else if (diff_cred < 0)
688 		racct_sub_cred_locked(p->p_ucred, resource, -diff_cred);
689 
690 	return (0);
691 }
692 
693 /*
694  * Set allocation of 'resource' to 'amount' for process 'p'.
695  * Return 0 if it's below limits, or errno, if it's not.
696  *
697  * Note that decreasing the allocation always returns 0,
698  * even if it's above the limit.
699  */
700 int
racct_set_unlocked(struct proc * p,int resource,uint64_t amount)701 racct_set_unlocked(struct proc *p, int resource, uint64_t amount)
702 {
703 	int error;
704 
705 	ASSERT_RACCT_ENABLED();
706 	PROC_LOCK(p);
707 	error = racct_set(p, resource, amount);
708 	PROC_UNLOCK(p);
709 	return (error);
710 }
711 
712 int
racct_set(struct proc * p,int resource,uint64_t amount)713 racct_set(struct proc *p, int resource, uint64_t amount)
714 {
715 	int error;
716 
717 	if (!racct_enable)
718 		return (0);
719 
720 	SDT_PROBE3(racct, , rusage, set__force, p, resource, amount);
721 
722 	RACCT_LOCK();
723 	error = racct_set_locked(p, resource, amount, 0);
724 	RACCT_UNLOCK();
725 	return (error);
726 }
727 
728 void
racct_set_force(struct proc * p,int resource,uint64_t amount)729 racct_set_force(struct proc *p, int resource, uint64_t amount)
730 {
731 
732 	if (!racct_enable)
733 		return;
734 
735 	SDT_PROBE3(racct, , rusage, set, p, resource, amount);
736 
737 	RACCT_LOCK();
738 	racct_set_locked(p, resource, amount, 1);
739 	RACCT_UNLOCK();
740 }
741 
742 /*
743  * Returns amount of 'resource' the process 'p' can keep allocated.
744  * Allocating more than that would be denied, unless the resource
745  * is marked undeniable.  Amount of already allocated resource does
746  * not matter.
747  */
748 uint64_t
racct_get_limit(struct proc * p,int resource)749 racct_get_limit(struct proc *p, int resource)
750 {
751 #ifdef RCTL
752 	uint64_t available;
753 
754 	if (!racct_enable)
755 		return (UINT64_MAX);
756 
757 	RACCT_LOCK();
758 	available = rctl_get_limit(p, resource);
759 	RACCT_UNLOCK();
760 
761 	return (available);
762 #else
763 
764 	return (UINT64_MAX);
765 #endif
766 }
767 
768 /*
769  * Returns amount of 'resource' the process 'p' can keep allocated.
770  * Allocating more than that would be denied, unless the resource
771  * is marked undeniable.  Amount of already allocated resource does
772  * matter.
773  */
774 uint64_t
racct_get_available(struct proc * p,int resource)775 racct_get_available(struct proc *p, int resource)
776 {
777 #ifdef RCTL
778 	uint64_t available;
779 
780 	if (!racct_enable)
781 		return (UINT64_MAX);
782 
783 	RACCT_LOCK();
784 	available = rctl_get_available(p, resource);
785 	RACCT_UNLOCK();
786 
787 	return (available);
788 #else
789 
790 	return (UINT64_MAX);
791 #endif
792 }
793 
794 /*
795  * Returns amount of the %cpu resource that process 'p' can add to its %cpu
796  * utilization.  Adding more than that would lead to the process being
797  * throttled.
798  */
799 static int64_t
racct_pcpu_available(struct proc * p)800 racct_pcpu_available(struct proc *p)
801 {
802 #ifdef RCTL
803 	uint64_t available;
804 
805 	ASSERT_RACCT_ENABLED();
806 
807 	RACCT_LOCK();
808 	available = rctl_pcpu_available(p);
809 	RACCT_UNLOCK();
810 
811 	return (available);
812 #else
813 
814 	return (INT64_MAX);
815 #endif
816 }
817 
818 /*
819  * Decrease allocation of 'resource' by 'amount' for process 'p'.
820  */
821 void
racct_sub(struct proc * p,int resource,uint64_t amount)822 racct_sub(struct proc *p, int resource, uint64_t amount)
823 {
824 
825 	if (!racct_enable)
826 		return;
827 
828 	SDT_PROBE3(racct, , rusage, sub, p, resource, amount);
829 
830 	/*
831 	 * We need proc lock to dereference p->p_ucred.
832 	 */
833 	PROC_LOCK_ASSERT(p, MA_OWNED);
834 	KASSERT(RACCT_CAN_DROP(resource),
835 	    ("%s: called for non-droppable resource %d", __func__, resource));
836 
837 	RACCT_LOCK();
838 	KASSERT(amount <= p->p_racct->r_resources[resource],
839 	    ("%s: freeing %ju of resource %d, which is more "
840 	     "than allocated %jd for %s (pid %d)", __func__, amount, resource,
841 	    (intmax_t)p->p_racct->r_resources[resource], p->p_comm, p->p_pid));
842 
843 	racct_adjust_resource(p->p_racct, resource, -amount);
844 	racct_sub_cred_locked(p->p_ucred, resource, amount);
845 	RACCT_UNLOCK();
846 }
847 
848 static void
racct_sub_cred_locked(struct ucred * cred,int resource,uint64_t amount)849 racct_sub_cred_locked(struct ucred *cred, int resource, uint64_t amount)
850 {
851 	struct prison *pr;
852 
853 	ASSERT_RACCT_ENABLED();
854 
855 	racct_adjust_resource(cred->cr_ruidinfo->ui_racct, resource, -amount);
856 	for (pr = cred->cr_prison; pr != NULL; pr = pr->pr_parent)
857 		racct_adjust_resource(pr->pr_prison_racct->prr_racct, resource,
858 		    -amount);
859 	racct_adjust_resource(cred->cr_loginclass->lc_racct, resource, -amount);
860 }
861 
862 /*
863  * Decrease allocation of 'resource' by 'amount' for credential 'cred'.
864  */
865 void
racct_sub_cred(struct ucred * cred,int resource,uint64_t amount)866 racct_sub_cred(struct ucred *cred, int resource, uint64_t amount)
867 {
868 
869 	if (!racct_enable)
870 		return;
871 
872 	SDT_PROBE3(racct, , rusage, sub__cred, cred, resource, amount);
873 
874 #ifdef notyet
875 	KASSERT(RACCT_CAN_DROP(resource),
876 	    ("%s: called for resource %d which can not drop", __func__,
877 	     resource));
878 #endif
879 
880 	RACCT_LOCK();
881 	racct_sub_cred_locked(cred, resource, amount);
882 	RACCT_UNLOCK();
883 }
884 
885 /*
886  * Inherit resource usage information from the parent process.
887  */
888 int
racct_proc_fork(struct proc * parent,struct proc * child)889 racct_proc_fork(struct proc *parent, struct proc *child)
890 {
891 	int i, error = 0;
892 
893 	if (!racct_enable)
894 		return (0);
895 
896 	/*
897 	 * Create racct for the child process.
898 	 */
899 	racct_create(&child->p_racct);
900 
901 	PROC_LOCK(parent);
902 	PROC_LOCK(child);
903 	RACCT_LOCK();
904 
905 #ifdef RCTL
906 	error = rctl_proc_fork(parent, child);
907 	if (error != 0)
908 		goto out;
909 #endif
910 
911 	/* Init process cpu time. */
912 	child->p_prev_runtime = 0;
913 	child->p_throttled = 0;
914 
915 	/*
916 	 * Inherit resource usage.
917 	 */
918 	for (i = 0; i <= RACCT_MAX; i++) {
919 		if (parent->p_racct->r_resources[i] == 0 ||
920 		    !RACCT_IS_INHERITABLE(i))
921 			continue;
922 
923 		error = racct_set_locked(child, i,
924 		    parent->p_racct->r_resources[i], 0);
925 		if (error != 0)
926 			goto out;
927 	}
928 
929 	error = racct_add_locked(child, RACCT_NPROC, 1, 0);
930 	error += racct_add_locked(child, RACCT_NTHR, 1, 0);
931 
932 out:
933 	RACCT_UNLOCK();
934 	PROC_UNLOCK(child);
935 	PROC_UNLOCK(parent);
936 
937 	if (error != 0)
938 		racct_proc_exit(child);
939 
940 	return (error);
941 }
942 
943 /*
944  * Called at the end of fork1(), to handle rules that require the process
945  * to be fully initialized.
946  */
947 void
racct_proc_fork_done(struct proc * child)948 racct_proc_fork_done(struct proc *child)
949 {
950 
951 	if (!racct_enable)
952 		return;
953 
954 #ifdef RCTL
955 	PROC_LOCK(child);
956 	RACCT_LOCK();
957 	rctl_enforce(child, RACCT_NPROC, 0);
958 	rctl_enforce(child, RACCT_NTHR, 0);
959 	RACCT_UNLOCK();
960 	PROC_UNLOCK(child);
961 #endif
962 }
963 
964 void
racct_proc_exit(struct proc * p)965 racct_proc_exit(struct proc *p)
966 {
967 	struct timeval wallclock;
968 	uint64_t pct_estimate, pct, runtime;
969 	int i;
970 
971 	if (!racct_enable)
972 		return;
973 
974 	PROC_LOCK(p);
975 	/*
976 	 * We don't need to calculate rux, proc_reap() has already done this.
977 	 */
978 	runtime = cputick2usec(p->p_rux.rux_runtime);
979 #ifdef notyet
980 	KASSERT(runtime >= p->p_prev_runtime, ("runtime < p_prev_runtime"));
981 #else
982 	if (runtime < p->p_prev_runtime)
983 		runtime = p->p_prev_runtime;
984 #endif
985 	microuptime(&wallclock);
986 	timevalsub(&wallclock, &p->p_stats->p_start);
987 	if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
988 		pct_estimate = (1000000 * runtime * 100) /
989 		    ((uint64_t)wallclock.tv_sec * 1000000 +
990 		    wallclock.tv_usec);
991 	} else
992 		pct_estimate = 0;
993 	pct = racct_getpcpu(p, pct_estimate);
994 
995 	RACCT_LOCK();
996 	racct_set_locked(p, RACCT_CPU, runtime, 0);
997 	racct_add_cred_locked(p->p_ucred, RACCT_PCTCPU, pct);
998 
999 	KASSERT(p->p_racct->r_resources[RACCT_RSS] == 0,
1000 	    ("process reaped with %ju allocated for RSS\n",
1001 	    p->p_racct->r_resources[RACCT_RSS]));
1002 	for (i = 0; i <= RACCT_MAX; i++) {
1003 		if (p->p_racct->r_resources[i] == 0)
1004 			continue;
1005 		if (!RACCT_IS_RECLAIMABLE(i))
1006 			continue;
1007 		racct_set_locked(p, i, 0, 0);
1008 	}
1009 
1010 #ifdef RCTL
1011 	rctl_racct_release(p->p_racct);
1012 #endif
1013 	racct_destroy_locked(&p->p_racct);
1014 	RACCT_UNLOCK();
1015 	PROC_UNLOCK(p);
1016 }
1017 
1018 /*
1019  * Called after credentials change, to move resource utilisation
1020  * between raccts.
1021  */
1022 void
racct_proc_ucred_changed(struct proc * p,struct ucred * oldcred,struct ucred * newcred)1023 racct_proc_ucred_changed(struct proc *p, struct ucred *oldcred,
1024     struct ucred *newcred)
1025 {
1026 	struct uidinfo *olduip, *newuip;
1027 	struct loginclass *oldlc, *newlc;
1028 	struct prison *oldpr, *newpr, *pr;
1029 
1030 	if (!racct_enable)
1031 		return;
1032 
1033 	PROC_LOCK_ASSERT(p, MA_OWNED);
1034 
1035 	newuip = newcred->cr_ruidinfo;
1036 	olduip = oldcred->cr_ruidinfo;
1037 	newlc = newcred->cr_loginclass;
1038 	oldlc = oldcred->cr_loginclass;
1039 	newpr = newcred->cr_prison;
1040 	oldpr = oldcred->cr_prison;
1041 
1042 	RACCT_LOCK();
1043 	if (newuip != olduip) {
1044 		racct_sub_racct(olduip->ui_racct, p->p_racct);
1045 		racct_add_racct(newuip->ui_racct, p->p_racct);
1046 	}
1047 	if (newlc != oldlc) {
1048 		racct_sub_racct(oldlc->lc_racct, p->p_racct);
1049 		racct_add_racct(newlc->lc_racct, p->p_racct);
1050 	}
1051 	if (newpr != oldpr) {
1052 		for (pr = oldpr; pr != NULL; pr = pr->pr_parent)
1053 			racct_sub_racct(pr->pr_prison_racct->prr_racct,
1054 			    p->p_racct);
1055 		for (pr = newpr; pr != NULL; pr = pr->pr_parent)
1056 			racct_add_racct(pr->pr_prison_racct->prr_racct,
1057 			    p->p_racct);
1058 	}
1059 	RACCT_UNLOCK();
1060 }
1061 
1062 void
racct_move(struct racct * dest,struct racct * src)1063 racct_move(struct racct *dest, struct racct *src)
1064 {
1065 
1066 	ASSERT_RACCT_ENABLED();
1067 
1068 	RACCT_LOCK();
1069 	racct_add_racct(dest, src);
1070 	racct_sub_racct(src, src);
1071 	RACCT_UNLOCK();
1072 }
1073 
1074 static void
ast_racct(struct thread * td,int tda __unused)1075 ast_racct(struct thread *td, int tda __unused)
1076 {
1077 	struct proc *p;
1078 
1079 	ASSERT_RACCT_ENABLED();
1080 
1081 	p = td->td_proc;
1082 	if (p->p_throttled == 0)
1083 		return;
1084 
1085 	PROC_LOCK(p);
1086 	while (p->p_throttled != 0) {
1087 		msleep(p->p_racct, &p->p_mtx, 0, "racct",
1088 		    p->p_throttled < 0 ? 0 : p->p_throttled);
1089 		if (p->p_throttled > 0)
1090 			p->p_throttled = 0;
1091 	}
1092 	PROC_UNLOCK(p);
1093 }
1094 
1095 /*
1096  * Make the process sleep in userret() for 'timeout' ticks.  Setting
1097  * timeout to -1 makes it sleep until woken up by racct_proc_wakeup().
1098  */
1099 void
racct_proc_throttle(struct proc * p,int timeout)1100 racct_proc_throttle(struct proc *p, int timeout)
1101 {
1102 	struct thread *td;
1103 #ifdef SMP
1104 	int cpuid;
1105 #endif
1106 
1107 	KASSERT(timeout != 0, ("timeout %d", timeout));
1108 	ASSERT_RACCT_ENABLED();
1109 	PROC_LOCK_ASSERT(p, MA_OWNED);
1110 
1111 	/*
1112 	 * Do not block kernel processes.  Also do not block processes with
1113 	 * low %cpu utilization to improve interactivity.
1114 	 */
1115 	if ((p->p_flag & (P_SYSTEM | P_KPROC)) != 0)
1116 		return;
1117 
1118 	if (p->p_throttled < 0 || (timeout > 0 && p->p_throttled > timeout))
1119 		return;
1120 
1121 	p->p_throttled = timeout;
1122 
1123 	FOREACH_THREAD_IN_PROC(p, td) {
1124 		thread_lock(td);
1125 		ast_sched_locked(td, TDA_RACCT);
1126 
1127 		switch (TD_GET_STATE(td)) {
1128 		case TDS_RUNQ:
1129 			/*
1130 			 * If the thread is on the scheduler run-queue, we can
1131 			 * not just remove it from there.  So we set the flag
1132 			 * TDA_SCHED for the thread, so that once it is
1133 			 * running, it is taken off the cpu as soon as possible.
1134 			 */
1135 			ast_sched_locked(td, TDA_SCHED);
1136 			break;
1137 		case TDS_RUNNING:
1138 			/*
1139 			 * If the thread is running, we request a context
1140 			 * switch for it by setting the TDA_SCHED flag.
1141 			 */
1142 			ast_sched_locked(td, TDA_SCHED);
1143 #ifdef SMP
1144 			cpuid = td->td_oncpu;
1145 			if ((cpuid != NOCPU) && (td != curthread))
1146 				ipi_cpu(cpuid, IPI_AST);
1147 #endif
1148 			break;
1149 		default:
1150 			break;
1151 		}
1152 		thread_unlock(td);
1153 	}
1154 }
1155 
1156 static void
racct_proc_wakeup(struct proc * p)1157 racct_proc_wakeup(struct proc *p)
1158 {
1159 
1160 	ASSERT_RACCT_ENABLED();
1161 
1162 	PROC_LOCK_ASSERT(p, MA_OWNED);
1163 
1164 	if (p->p_throttled != 0) {
1165 		p->p_throttled = 0;
1166 		wakeup(p->p_racct);
1167 	}
1168 }
1169 
1170 static void
racct_decay_callback(struct racct * racct,void * dummy1,void * dummy2)1171 racct_decay_callback(struct racct *racct, void *dummy1, void *dummy2)
1172 {
1173 	int64_t r_old, r_new;
1174 
1175 	ASSERT_RACCT_ENABLED();
1176 	RACCT_LOCK_ASSERT();
1177 
1178 #ifdef RCTL
1179 	rctl_throttle_decay(racct, RACCT_READBPS);
1180 	rctl_throttle_decay(racct, RACCT_WRITEBPS);
1181 	rctl_throttle_decay(racct, RACCT_READIOPS);
1182 	rctl_throttle_decay(racct, RACCT_WRITEIOPS);
1183 #endif
1184 
1185 	r_old = racct->r_resources[RACCT_PCTCPU];
1186 
1187 	/* If there is nothing to decay, just exit. */
1188 	if (r_old <= 0)
1189 		return;
1190 
1191 	r_new = r_old * RACCT_DECAY_FACTOR / FSCALE;
1192 	racct->r_resources[RACCT_PCTCPU] = r_new;
1193 }
1194 
1195 static void
racct_decay_pre(void)1196 racct_decay_pre(void)
1197 {
1198 
1199 	RACCT_LOCK();
1200 }
1201 
1202 static void
racct_decay_post(void)1203 racct_decay_post(void)
1204 {
1205 
1206 	RACCT_UNLOCK();
1207 }
1208 
1209 static void
racct_decay(void)1210 racct_decay(void)
1211 {
1212 
1213 	ASSERT_RACCT_ENABLED();
1214 
1215 	ui_racct_foreach(racct_decay_callback, racct_decay_pre,
1216 	    racct_decay_post, NULL, NULL);
1217 	loginclass_racct_foreach(racct_decay_callback, racct_decay_pre,
1218 	    racct_decay_post, NULL, NULL);
1219 	prison_racct_foreach(racct_decay_callback, racct_decay_pre,
1220 	    racct_decay_post, NULL, NULL);
1221 }
1222 
1223 static void
racctd(void)1224 racctd(void)
1225 {
1226 	struct thread *td;
1227 	struct proc *p;
1228 	struct timeval wallclock;
1229 	uint64_t pct, pct_estimate, runtime;
1230 
1231 	ASSERT_RACCT_ENABLED();
1232 
1233 	for (;;) {
1234 		racct_decay();
1235 
1236 		sx_slock(&allproc_lock);
1237 
1238 		FOREACH_PROC_IN_SYSTEM(p) {
1239 			PROC_LOCK(p);
1240 			if (p->p_state != PRS_NORMAL ||
1241 			    (p->p_flag & P_IDLEPROC) != 0) {
1242 				if (p->p_state == PRS_ZOMBIE)
1243 					racct_set(p, RACCT_PCTCPU, 0);
1244 				PROC_UNLOCK(p);
1245 				continue;
1246 			}
1247 
1248 			microuptime(&wallclock);
1249 			timevalsub(&wallclock, &p->p_stats->p_start);
1250 			PROC_STATLOCK(p);
1251 			FOREACH_THREAD_IN_PROC(p, td)
1252 				ruxagg(p, td);
1253 			runtime = cputick2usec(p->p_rux.rux_runtime);
1254 			PROC_STATUNLOCK(p);
1255 #ifdef notyet
1256 			KASSERT(runtime >= p->p_prev_runtime,
1257 			    ("runtime < p_prev_runtime"));
1258 #else
1259 			if (runtime < p->p_prev_runtime)
1260 				runtime = p->p_prev_runtime;
1261 #endif
1262 			p->p_prev_runtime = runtime;
1263 			if (wallclock.tv_sec > 0 || wallclock.tv_usec > 0) {
1264 				pct_estimate = (1000000 * runtime * 100) /
1265 				    ((uint64_t)wallclock.tv_sec * 1000000 +
1266 				    wallclock.tv_usec);
1267 			} else
1268 				pct_estimate = 0;
1269 			pct = racct_getpcpu(p, pct_estimate);
1270 			RACCT_LOCK();
1271 #ifdef RCTL
1272 			rctl_throttle_decay(p->p_racct, RACCT_READBPS);
1273 			rctl_throttle_decay(p->p_racct, RACCT_WRITEBPS);
1274 			rctl_throttle_decay(p->p_racct, RACCT_READIOPS);
1275 			rctl_throttle_decay(p->p_racct, RACCT_WRITEIOPS);
1276 #endif
1277 			racct_set_locked(p, RACCT_PCTCPU, pct, 1);
1278 			racct_set_locked(p, RACCT_CPU, runtime, 0);
1279 			racct_set_locked(p, RACCT_WALLCLOCK,
1280 			    (uint64_t)wallclock.tv_sec * 1000000 +
1281 			    wallclock.tv_usec, 0);
1282 			RACCT_UNLOCK();
1283 			PROC_UNLOCK(p);
1284 		}
1285 
1286 		/*
1287 		 * To ensure that processes are throttled in a fair way, we need
1288 		 * to iterate over all processes again and check the limits
1289 		 * for %cpu resource only after ucred racct containers have been
1290 		 * properly filled.
1291 		 */
1292 		FOREACH_PROC_IN_SYSTEM(p) {
1293 			PROC_LOCK(p);
1294 			if (p->p_state != PRS_NORMAL) {
1295 				PROC_UNLOCK(p);
1296 				continue;
1297 			}
1298 
1299 			if (racct_pcpu_available(p) <= 0) {
1300 				if (p->p_racct->r_resources[RACCT_PCTCPU] >
1301 				    pcpu_threshold)
1302 					racct_proc_throttle(p, -1);
1303 			} else if (p->p_throttled == -1) {
1304 				racct_proc_wakeup(p);
1305 			}
1306 			PROC_UNLOCK(p);
1307 		}
1308 		sx_sunlock(&allproc_lock);
1309 		pause("-", hz);
1310 	}
1311 }
1312 
1313 static struct kproc_desc racctd_kp = {
1314 	"racctd",
1315 	racctd,
1316 	NULL
1317 };
1318 
1319 static void
racctd_init(void)1320 racctd_init(void)
1321 {
1322 	if (!racct_enable)
1323 		return;
1324 
1325 	kproc_start(&racctd_kp);
1326 }
1327 SYSINIT(racctd, SI_SUB_RACCTD, SI_ORDER_FIRST, racctd_init, NULL);
1328 
1329 static void
racct_init(void)1330 racct_init(void)
1331 {
1332 	if (!racct_enable)
1333 		return;
1334 
1335 	racct_zone = uma_zcreate("racct", sizeof(struct racct),
1336 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
1337 	ast_register(TDA_RACCT, ASTR_ASTF_REQUIRED, 0, ast_racct);
1338 
1339 	/*
1340 	 * XXX: Move this somewhere.
1341 	 */
1342 	prison0.pr_prison_racct = prison_racct_find("0");
1343 }
1344 SYSINIT(racct, SI_SUB_RACCT, SI_ORDER_FIRST, racct_init, NULL);
1345